CNN v2: Fix weight buffer offset bug

Root cause: Binary format is [header:20B][layer_info:20B×N][weights].
Both cnn_test and CNNv2Effect uploaded entire file to weights_buffer,
but shader reads weights_buffer[0] expecting first weight, not header.

Fix: Skip header + layer_info when uploading to GPU buffer.
- cnn_test.cc: Calculate weights_offset, upload only weights section
- cnn_v2_effect.cc: Same fix for runtime effect

Before: layer_0 output showed [R, uv_x, uv_y, black] (wrong channels)
After: layer_0 output shows [R, G, B, D] (correct identity mapping)

Tests: 34/36 passing (2 unrelated failures)

2 files changed, 18 insertions, 9 deletions

diff --git a/src/gpu/effects/cnn_v2_effect.cc b/src/gpu/effects/cnn_v2_effect.cc
index 3985723..366a232 100644
--- a/src/gpu/effects/cnn_v2_effect.cc
+++ b/src/gpu/effects/cnn_v2_effect.cc
@@ -111,17 +111,21 @@ void CNNv2Effect::load_weights() {
     layer_info_.push_back(info);
   }
 
-  // Create GPU storage buffer for weights
-  // Buffer contains: header + layer info + packed f16 weights (as u32)
+  // Create GPU storage buffer for weights (skip header + layer info, upload only weights)
+  size_t header_size = 20;  // 5 u32
+  size_t layer_info_size = 20 * num_layers;  // 5 u32 per layer
+  size_t weights_offset = header_size + layer_info_size;
+  size_t weights_only_size = weights_size - weights_offset;
+
   WGPUBufferDescriptor buffer_desc = {};
-  buffer_desc.size = weights_size;
+  buffer_desc.size = weights_only_size;
   buffer_desc.usage = WGPUBufferUsage_Storage | WGPUBufferUsage_CopyDst;
   buffer_desc.mappedAtCreation = false;
 
   weights_buffer_ = wgpuDeviceCreateBuffer(ctx_.device, &buffer_desc);
 
-  // Upload weights data
-  wgpuQueueWriteBuffer(ctx_.queue, weights_buffer_, 0, weights_data, weights_size);
+  // Upload only weights (skip header + layer info)
+  wgpuQueueWriteBuffer(ctx_.queue, weights_buffer_, 0, weights_data + weights_offset, weights_only_size);
 
   // Create uniform buffers for layer params (one per layer)
   for (uint32_t i = 0; i < num_layers; ++i) {
diff --git a/tools/cnn_test.cc b/tools/cnn_test.cc
index 4599512..c504c3d 100644
--- a/tools/cnn_test.cc
+++ b/tools/cnn_test.cc
@@ -684,15 +684,20 @@ static bool process_cnn_v2(WGPUDevice device, WGPUQueue queue,
            info.out_channels, info.weight_count);
   }
 
-  // Create weights storage buffer
+  // Create weights storage buffer (skip header + layer info, upload only weights)
+  size_t header_size = 20;  // 5 u32
+  size_t layer_info_size = 20 * layer_info.size();  // 5 u32 per layer
+  size_t weights_offset = header_size + layer_info_size;
+  size_t weights_only_size = weights_size - weights_offset;
+
   WGPUBufferDescriptor weights_buffer_desc = {};
-  weights_buffer_desc.size = weights_size;
+  weights_buffer_desc.size = weights_only_size;
   weights_buffer_desc.usage = WGPUBufferUsage_Storage | WGPUBufferUsage_CopyDst;
   weights_buffer_desc.mappedAtCreation = false;
 
   WGPUBuffer weights_buffer =
       wgpuDeviceCreateBuffer(device, &weights_buffer_desc);
-  wgpuQueueWriteBuffer(queue, weights_buffer, 0, weights_data, weights_size);
+  wgpuQueueWriteBuffer(queue, weights_buffer, 0, weights_data + weights_offset, weights_only_size);
 
   // Create input view
   const WGPUTextureViewDescriptor view_desc = {
@@ -1051,7 +1056,7 @@ static bool process_cnn_v2(WGPUDevice device, WGPUQueue queue,
 
     layer_bg_entries[3].binding = 3;
     layer_bg_entries[3].buffer = weights_buffer;
-    layer_bg_entries[3].size = weights_size;
+    layer_bg_entries[3].size = weights_only_size;
 
     layer_bg_entries[4].binding = 4;
     layer_bg_entries[4].buffer = layer_params_buffers[i];